模型與開發者 API
Mistral OCR 4.1 Fixes Bounding Box Drift and Quote Rewriting; `latest` Alias Switches Immediately
Mistral OCR 4.1 focuses on six categories of document-parsing errors, including bounding box drift, missing blocks, merged references, and right-to-left tables. `mistral-ocr-latest` now points to the new version, so production pipelines relying on the floating alias may see output changes without code changes or version pinning.
Mistral released OCR 4.1 on August 13. Rather than adding new API interfaces, this release fixes six categories of model behavior that directly affect document extraction, RAG citations, and structured data pipelines. The new version makes paragraph bounding boxes align more closely with the actual content, preventing image boxes from becoming nested or coordinates from progressively drifting. On reference pages, it now returns a separate bounding box for each citation instead of collapsing an entire list into a single block. Mistral also says the model misses fewer blocks on complex pages, recognizes more checkbox styles, and parses tables in right-to-left languages more effectively.
Another fix that may seem minor but could break downstream applications is that OCR no longer changes double quotation marks in documents to single quotation marks on its own. When results feed into code extraction, contract comparison, JSON post-processing, or verbatim quotation systems, punctuation normalization can distort content and even cause a parser to produce different semantics. The OCR 4 response format remains centered on per-page Markdown, images, dimensions, and optional structured blocks, so existing clients generally do not need to change their schema.
The version alias requires immediate attention. Mistral says `mistral-ocr-latest` now points directly to 4.1, so requests using that name automatically receive the new version. This makes it easy to pick up fixes, but it also means that the text, number of bounding boxes, and coordinates produced for the same input may change without any new code being deployed. Engineering teams should build a golden set from their own scanned documents and specifically retest citation localization, checkboxes, RTL tables, and quote-sensitive fields, while retaining the exact model version and raw responses. The announcement provides no new independent benchmarks, per-language scores, or error rates, so the six improvements remain primarily vendor-reported. Pipelines that require reproducible stability should not rely solely on a floating alias.